Online Multiple Object Tracking Based on State Prediction and Motion Structure

Tracking-by-detection has emerged as a research focus of multiple object tracking(MOT)in recent years.Most online MOT methods took only two adjacent frames into consideration when dealing with the task of data association,which led to unstable performance with false positives and problem of trajectory fragments.In this paper,the proposed scheme includes two stages,tracking frame by frame and trajectory recovery.In the tracking stage,motion structure of objects and multiple feature fusion are leveraged to improve the accuracy of object matching.Meanwhile leveraging motion structure will also maintain robustness to camera movements.In the recovery stage,a state prediction method is leveraged to estimate the potential states of the non-associated objects,which are used to re-associate with detections.This stage aims at solving the problem of trajectory fragments.The proposed scheme was evaluated on the MOTChallenge 2015 benchmarks.Experimental results show its efficiency and robustness.